Fixed the rounding bug in Optimized Composite Ops

The vector and single-pixel versions of the op must generate absolutely
the same pixel values and with absolutely the same rounding rules.

This patch implements not-the-best way of rounding, because Vc doesn't
support single instruction round-and-convert-to-int opcodes (which are
present in the Intel architecture). Ideally, we should implement
Vc::iRount() method to do this.

Right now, using a separate-opcode rounding makes CompositeOver about
5-10% slower.

TODO: Implement Vc::iRound()